Using U-Net to Detect Landslides
In Spring 2021, I began working with Dr. James Kirkpatrick at McGill University to implement a customized deep learning model for identifying landslides in geospatial imagery. Originally, our goal was to predict landslides in a spatiotemporal context. We very quickly learned there was a severe lack of open, well-labeled data available for this project. Instead, we decided to focus on building a model that could create the kind of data we lacked.
Based on the U-Net image segmentation model originally developed for biomedical image segmentation by Ronneberger et al. (2015), we developed a custom model that could identify landslides in satellite imagery. The California Geological Survey (CGS) was kind enough to provide us with a small portion of their beta dataset of landslide scarp and deposit data in Los Briones, CA. The most recent iteration of the model was able to identify landslides with 95.3% accuracy and a loss of 0.19, but we are still facing challenges with the F1 Score, recall, and precision. In the near future, we hope to improve our model by using a larger dataset and by implementing a more robust data augmentation pipeline.
The project has been on and off at times and is currently on hold while I work on a related project.